Using the learning classifier system for robot navigation

Exhibitor: Chamindra Rajakaruna

Supervisor: Tom Downs

Research Group: Complex and Intelligent Systems

Industry Sector: Manufacturing and Industrial Applications

The classical robotics problem involved programmers’ hand coding actions for robot operations. This is often a tedious and time-consuming procedure as designers are required to understand the robot’s operational environment completely and account for every possible situation that the robot may encounter. This is a virtually impossible task in many real world applications of robots and as a result saw the robots often designated to carry out repetitive and simple tasks on factory assembly lines and other industrial applications. Developments in areas such as machine learning, neural networks and learning classifier systems have given rise to the field of evolutionary robotics. This allows designers to develop robots for operation in unknown or dynamic environments. By implementing “self learning” mechanisms, they are able to construct robots with minimal initial information, for example, they may now be able to develop robots for navigating in environments for which there is very limited information available such as Mars or the deep sea. This adaptability and flexibility is one of the defining features in the area of evolutionary robotics. Robots are no longer restricted to simple repetitive task but are now able to explore new and uncharted horizons.

The primary scope of this thesis project focuses on the development of obstacle avoidance in an autonomous mobile robot. The primary learning mode utilised has been the use of a Learning classifier, which provides the robot with a self-learning ability. The leaning classifier has minimised the involvement of the programmer in the teaching process. The programmer simply outlines the learning algorithms and the robot begins the learning process by simply interacting with its environment.

Much of the initial groundwork had already been laid in the form of a simulated robot navigation program known as the rudimentary classifier system. I have focused on extending and improving this simulation by improving the learning and reinforcement modules within the robot. I have also laid the groundwork required for the transference of the simulated robot to a physical robot able to learn and interact with the real world environment.

Thesis Document (PDF)